MolGan

MolGAN (Molecular Generative Adversarial Network), a deep learning model designed to create new, chemically valid molecular structures.

Output

1. Using pure RL (lamda = 0)

2. Half wgan and RL (lambda = 0.5)

3. Pure wgan (lambda = 1)

Architecture

The architecture consists of 3 main sections: a generator, a discriminator, and a reward network.

The generator takes a sample (z) from a standard normal distribution to generate a graph using an MLP (this limits the network to a fixed maximum size) to generate the graph at once. Specifically a dense adjacency tensor A (bond types) and an annotation matrix X (atom types) are produced. Since these are probabilities, a discrete, sparse x and a are generated through categorical sampling.

The discriminator and reward network have the same architectures and receive graphs as inputs. A Relational-GCN and MLPs are used to produce the singular output

Training Steps

1. Clone the Repository
   Clone the MolGAN implementation repository to your local machine:

   git clone https://github.com/kfzyqin/Implementation-MolGAN-PyTorch.git
   cd Implementation-MolGAN-PyTorch

2. Download Molecular Metrics Models
   Run the provided script to download pre-trained models for evaluating molecular metrics (e.g., validity, uniqueness, novelty):

   sh download_dataset.sh

   This script downloads necessary utilities defined in utils.py.

3. Generate the Dataset
   Process the QM9 dataset to create a sparse molecular dataset compatible with MolGAN:

   python sparse_molecular_dataset.py

   Note: Uncomment the last few lines in sparse_molecular_dataset.py to enable dataset generation.

4. Train the Model
   Execute the training script to start training MolGAN:

   python train.py

   • Hyperparameters (e.g., learning rate, batch size, λ) can be adjusted in the train.py or Solve module for experimentation.
   • Monitor training progress through loss metrics and generated molecule quality.

References:

1. Molgan paper
2. Implementation-MolGan-Pytorch